前沿 AI 安全监管与跨部门事故报告框架

10 Dr Charlene Chen asked the Minister for Digital Development and Information (a) whether the Singapore AI Safety Institute's ability to independently evaluate advanced AI models has been constrained by insufficient access from developers; (b) how risks are assessed where such access is unavailable; and (c) whether independent safety testing will be required for AI systems assessed to pose sufficiently high risks.

11 Dr Charlene Chen asked the Minister for Digital Development and Information (a) whether the Government has assessed whether existing sectoral incident-reporting requirements adequately capture serious incidents involving AI systems; and (b) where gaps remain, whether a cross-sector reporting framework for serious AI incidents will be considered.

12 Mr Alex Yam asked the Minister for Digital Development and Information in light of the recent unauthorised access of the Australian government's systems by an autonomous AI agent (a) whether Singapore has assessed the resilience of Government digital systems against similar agentic AI behaviour; and (b) whether existing cybersecurity and incident-reporting frameworks adequately address autonomous actions by AI systems that were not specifically directed by human operators.

13 Dr Choo Pei Ling asked the Minister for Digital Development and Information (a) whether the Government will introduce mandatory safety requirements for frontier AI models deployed in Singapore, including pre-deployment risk assessments, independent safety testing, safeguards against misuse and reporting of serious incidents; and (b) if so, how will these requirements be enforced and updated as AI capabilities evolve.

14 Mr Yip Hon Weng asked the Minister for Digital Development and Information following Singapore’s support for international oversight of frontier AI (a) how the Ministry determines whether overseas models are sufficiently safe for deployment in essential services when independent testing access is unavailable; and (b) which considerations would justify requiring verifiable safety evidence before deployment, rather than relying on developers' assurances.

15 Dr Choo Pei Ling asked the Minister for Digital Development and Information (a) whether the Government has set thresholds or indicators for when rapid AI model advancements will require additional safeguards beyond existing cybersecurity requirements; and (b) whether Singapore has contingency arrangements with foreign governments, major AI developers, and international cybersecurity bodies for cross-border AI-enabled cyber incidents.

The Senior Minister of State for Digital Development and Information (Mr Tan Kiat How) (for the Minister for Digital Development and Information) : Sir, for the Question Nos 10 to 15 on the Order Paper today, these questions have been addressed in the reply to Question Nos 45 to 51 and Question No 33 for written reply at yesterday's Parliamentary Sitting. [ Please refer to " Assessment of Singapore's Digital Systems' Resilience Against Agentic AI Behaviours ", Official Report, 6 October 2026, Vol 96, Issue 38, Written Answers to Questions for Oral Answer not Answered by End of Question Time section. ]

Mr Speaker : Dr Chen.

Dr Charlene Chen (Tampines) : Thank you, Speaker. In her reply yesterday, the Minister mentioned the importance of learning from both incidents and near misses. Could the Senior Minister of State clarify what threshold would make an artificial intelligence (AI) incident reportable, and whether significant near misses should also be reported even when human intervention prevents actual harm?

My second supplementary question is, much of the Ministry's response has focused on Government systems and essential services. But AI systems are also increasingly being deployed by private organisations. For an ordinary citizen who may be seriously affected by an unintended action of an AI system, for example, involving their personal or financial information, what protections and avenues for redress are in place? And would they have a right to be informed that an AI-related incident has affected them?

My last supplementary question. If an AI system already deployed in Singapore is found to pose an immediate and serious risk to the public, do our regulators currently have the power to require its use to be suspended or restricted while the risk is investigated, including if the underlying model comes from overseas?

Mr Tan Kiat How : Sir, I thank the Member for her three supplementary questions. They are broadly aligned to the same theme, around how we are thinking about risks and what the responses are. So, let me take all her three supplementary questions together.

First, let me say that we recognise that AI, and increasingly, agentic and autonomous AI systems, presents new opportunities and risks. The difference for agentic systems is that these systems can execute actions, interact with external systems and adapt their behaviour to reduce human oversight. So, our response focuses on two areas.

First, we should recognise that many risks arising from agentic AI and some of the autonomous systems that the Member has alluded to are extensions of existing challenges. So, whether the impact is on the individual or the impact is on a critical information infrastructure system or another essential digital service, these are extensions of the same challenges that we face today.

Therefore, existing guidelines and regulations, including risk management processes and oversight also apply to agentic AI systems. The principle of maintaining human accountability and putting in place sufficient controls and guardrails also applies. In addition, our longstanding efforts into strengthening data protection, cybersecurity and resilience help protect the systems that AI agents interact with.

Of course, we are reviewing how to adapt and strengthen these frameworks and systems to account for the increased autonomy of agentic systems, while also monitoring international developments in this emerging field.

So, the first point that I would like to emphasise is that these risks are already with us. Yes, there might be some new dimensions to it, but these are existing risks, and existing frameworks apply and we are reviewing how to strengthen them.

Second, we have to also develop our own internal capabilities. We are developing capabilities to go beyond existing frameworks and approaches. We are running trials and experiments to stay abreast of the evolving technology and to understand what it takes to implement agentic systems responsibly. We certainly do so within the Government – the Government Technology Agency of Singapore (GovTech) does it and many other agencies within the Government are already doing it with the right safeguards and doing so responsibly.

I encourage all private sector organisations, even those who are not running critical information infrastructure or essential digital services, to also take appropriate care and responsibility when deploying AI agents and systems within their operations. But we are developing capabilities and this is a new area, emerging area, and rapidly evolving. We will continue to keep track of it and update our frameworks and regulations where needed.

Mr Speaker : Mr Alex Yam.

Mr Alex Yam (Marsiling-Yew Tee) : Thank you, Speaker. I thank the Senior Minister of State for his response and the Minister's reply yesterday. I want to focus on a particular aspect of the Open AI case in Australia, which I think is particularly important for us.

The Australian government was made aware of the unauthorised access only three months after the incident, because the developer eventually identified and disclosed it to the Australian government. So, in our local example, our legislation requires CII and critical infrastructure companies to report incidents. But is there a clear obligation today for developers, especially those who are overseas, to notify our authorities promptly and early so that we can take the appropriate measures?

The second question is about testing. I am quite certain that our local agencies, including the Cyber Security Agency and our agencies under the Ministry of Digital Development and Information do quite robust testings of systems and agents, and develop and models. While I support stronger independent testing, the examples that we have seen recently were within bounds of the developers' work and even those testing have gone out of hand. There is nothing nefarious that is going on. It is just a developer testing their model and those have also gone out of hand. So, what other safeguards can we put in place to ensure that even in our own testing capabilities, to discover dangers that we do not in itself create risk for other third parties as well?

Mr Tan Kiat How : Sir, let me address the Member's question in turn.

The first is around the ongoing matter; it is still a live issue. I have read publicly that the frontier AI company the Member mentioned is in correspondence with the Australian government. So, it is a live issue and we are monitoring it. I would say that we certainly welcome frontier AI companies to take greater responsibility in understanding and managing the risk arising from the systems that they develop, and this includes identifying emerging capabilities and risk early, and putting in place safeguards before these systems are widely deployed.

Where these systems or the AI frontier models' capabilities have caused harm or have caused impact on others, we certainly welcome that they take responsibility and to rectify the issue as soon as possible.

At the same time, I think the landscape has changed drastically over the last few months. We look at the threat landscape arrayed against us – and we have different cyber risk, digital risk that are arrayed against us – from cyber criminals, cyber hackers, advanced persistent threat (APT), which are state-sponsored, state-supported actors, or just individuals who, for the thrill of it, adopt digital solutions to hack into another system.

The risk arrayed against any organisation, any economy is very different compared to a few months ago. It is no longer just about the most frontier AI models being used to detect vulnerabilities or being able to exploit vulnerabilities by autonomously chaining actions together, but there are many open AI models out there online that is freely downloadable. Bad actors can use those capabilities to create cyber threats against organisations or for criminal enterprise.

So, the threat landscape is very different. It is not just about developers of frontier AI models being a risk, but there are many, many actors out there who can download open-weight models, adapt them, create their own harness, create their own contexts and use them for bad purposes. So, we cannot just wait for other people to tell us that they are hacking us. It is the onus on Singapore and our organisations, especially those running critical information infrastructure, to take the necessary safeguards.

There are many things that we can do, many things that the organisations running these essential services can do.

Firstly, making sure basic cyber hygiene is put in place. I often use an analogy: when you build information technology systems, it is like building new rooms and new wings or new annexes to your building. Sometimes, you may not know whether you have locked all your doors or secured all your windows. The agents or bad actors exploit the same vulnerabilities – an unlocked door or the unsecured window – to enter our systems, whether through AI means or more traditional conventional means.

So, first and foremost, take the necessary precaution: improve cyber hygiene, close down our vulnerabilities and continue an active monitoring of anomalous behaviour within our system. These are things – basic hygiene – things that we can do.

Second, even as bad actors use AI capabilities, we should as well be able to deploy those AI capabilities within our own system to make sure that we detect our vulnerabilities earlier, and we should have an advantage because we know our own system. We know the layout, the blueprint of our building. We know where the annexes are. We know where the critical services are in that building. So, we use AI to our own advantage. And that is the second thing I think we should do.

Thirdly, more importantly, is that we should not allow fear or anxiety about those risks to prevent us from making full use of the technology. So, we should use the technology to create value, new products, new services, create new opportunities for our people. But you do so responsibly, putting the right safeguards, learning along the way and adapting.

To the Member’s point, these are practical things we can do, without waiting for someone else to tell us that they have been hacking us.

Mr Speaker : Dr Choo Pei Ling.

Dr Choo Pei Ling (Chua Chu Kang) : Thank you, Speaker, and I thank the Senior Minister of State for his clarifications and response. I have a few supplementary questions.

One, based on what the Senior Minister of State has shared, could he further elaborate what would trigger the Government to move from guidance to mandatory safeguards before deployment of a frontier AI system? For example, when it demonstrates substantially greater autonomy, the ability to act across external systems or risks that existing safeguards cannot adequately contain.

The second supplementary question, if such a trigger is reached, how quickly can these stronger safeguards or in the Senior Minister of State’s previous terms, “AI capabilities” be applied, including where the model is developed overseas?

And lastly, have we had cross-border contingency arrangements that have been tested through joint exercises with foreign counterparts and major AI developers, and do they establish clear responsibilities for early warning, information sharing and containment?

Mr Tan Kiat How : Sir, let me address Dr Choo’s questions in turn. She asked whether there are any thresholds that turn the guidelines into mandatory requirements. This is something that we have to continue monitoring and reviewing. But let me say that the risk that we are dealing with and the steps that we can take to mitigate those risks are actually not new risks. These are existing risks – whether there is a cyberattack on our critical information infrastructure, or there is a service disruption caused by using AI bots or agents within an essential digital service, for example, in the financial services or other utilities. These are existing risks. From a cybersecurity point of view, the sectoral regulators would have to work closely with the operators of those critical information infrastructure services to look at the safeguards that need to be put in place.

So, whether it is a voluntary safeguard, for example, on the use of agentic AI in certain use cases or whatever kind of safeguards that needs to be put in place for certain deployment, those will have to be discussed between the sectoral regulators and the essential services operators.

And certainly, we do so for different services within the Public Service. For example, within the Government, GovTech works very closely with the different service owners of the public services to see how we use AI responsibly.

On international engagements, this is a very nascent, emerging area. We read the headlines. We welcome, for example, greater conversation and dialogue on AI safety. For Singapore, we try to play a constructive role, for instance through technical and scientific conferences, which we host. And there are a number of these conferences that we hold in Singapore with luminaries and very technically credible, frontier professionals in this space who are coming to Singapore, to not just have a conversation among themselves, but also interact with us and share their perspectives with us.

We have been also working very closely with different AI safety institutes. Our AI Safety Institute in Singapore has been working very closely with other international counterparts to share best practices.

So, we are deeply plugged into the scientific and engineering conversations that are happening. But this is a new area, something that is still being discussed. More recently, at multilateral fora, like the United Nations, we welcome more like-minded countries to work together. Singapore has recently signed a statement of support with like-minded countries to have greater international collaboration on AI safety. This is something we will continue to work on very closely with our international counterparts.

Mr Speaker : Last supplementary question, Mr Yip Hon Weng.

Mr Yip Hon Weng (Yio Chu Kang) : Thank you, Mr Speaker. I thank the Senior Minister of State for his response. Just a very short supplementary question. It is really to respond to the issue of the speed of development of AI.

As frontier models and their capabilities change rapidly after deployment, will the Ministry consider requiring periodic re-evaluation and clear incident reporting arrangements for essential service use, rather than treating pre-deployment assurance as a once-off exercise?

I ask this because this would help preserve public trust when systems evolve rapidly.

Mr Tan Kiat How : Sir, I thank the Member for his question. The existing regulations and rules do have incident reporting frameworks. We will continue to review those frameworks to see whether they need to be updated, given the different threat landscape and operating environment. But those incident reporting frameworks thresholds are already in place. For example, under the Personal Data Protection Act.

Regardless whether there are AI tools being used, organisations remain subject to those existing obligations. This includes putting in place reasonable security arrangements to protect personal data, and appropriate policies and practices to meet their data protection obligations.

Some Members will have seen in the news that there was a recent incident that occurred due to personal use of a GenAI tool by an organisation's employee who developed an email distribution script for sending marketing emails, but did so incorrectly, resulting in personal data being revealed.

This case underscores the need for organisations to manage the risk associated with widespread personal use of AI tools, which is another risk that many organisations face – so, not just frontier AI models, but what, in the industry, they call “shadow AI”. Which means that organisations may not know that your employee may be bringing AI tools into your organisation’s system and these are not sanctioned or approved by the organisation. But it is so easily accessible; you download it on your phone, use it for yourself in the workplace, without the necessary safeguards.

So, as AI tools become more accessible, employees may use them on their own initiative, which is something commendable – we should not discourage such productive use of AI. But organisations need to be aware of how these tools are being used, and put appropriate policies and safeguards in place.

So, those reporting thresholds and frameworks are in place. Importantly, it is for organisations to build a culture while encouraging employees to use AI tools to have the right safeguards and to be mindful that new risks are introduced because it is so much easier to bring AI tools into any organisation today.